Features / NCAAB

---

I check the BXS rankings every morning with coffee. Through November and December, the top 10 barely moved. Arizona, Michigan, and Houston sat in a comfortable cluster at the top, and the model felt almost bored. Then January hit, conference play kicked in, and the rankings started reshuffling faster than I could keep up. Duke gained 73 BXS points. Oklahoma lost eight games in a row. A mid-major I'd barely heard of climbed 64 points in 10 games.

This is NCAAB Wrapped: a monthly look at where the BXS model's numbers moved and what that movement actually means. January is the first month where the data gets real. Of the 1,558 games played, 1,549 were conference games. That's 99.4%. The cupcake era is over. Teams are finally playing opponents the model respects. And the results are telling.

The January Power Rankings

After a full month of conference play, the model's top 15 looks noticeably different from the AP poll:

Source: Boxscorus NCAAB BXS model, end of January 2026.

Five Big Ten teams in the top 11. Three Big 12 teams in the top 6. The SEC has Florida and Vanderbilt. The model doesn't care about brand names or preseason rankings. It cares about who beat whom and by how much.

The Biggest Risers: Who's Heating Up

Gaining BXS in conference play is harder than in non-conference. You're playing opponents rated closer to your own level, so each game's expected outcome is tighter. The model's margin-of-victory multiplier, (margin + 3)^0.8 / (7.5 + 0.006 * bxs_diff), adds a wrinkle: the denominator grows with the pre-game BXS gap, so blowing out a team you're already heavily favored over earns less credit than blowing out a team close to your rating. In conference play, where the gaps are smaller, big wins count more.

These teams gained anyway:

Source: Boxscorus NCAAB BXS model, January 2026 game data.

UIC gained 64 BXS points in 10 Missouri Valley Conference games. To put that in context, 64 points is roughly the gap between the 8th and 15th best teams in the country right now. That's not just winning. That's winning by margins the model didn't predict, game after game.

Saint Louis went 8-0 in the Atlantic-10, including a 97-62 demolition of St. Bonaventure and a 102-71 blowout of Dayton. The model's opponent quality factor already discounts A-10 wins relative to Big 12 or SEC wins (beating a 1500-rated team earns a K-factor of 1.0, while beating an 1800-rated team earns roughly 1.2x). That Saint Louis still gained +58 after those discounts tells you the margins were genuinely impressive.

Virginia's +56 is the power-conference riser on this list. They went 7-1 in the ACC, with their only loss at North Carolina (80-85). Wins over NC State (76-61), California (84-60), and Stanford (70-55) show a team beating ACC opponents by comfortable margins. At 1797, Virginia has quietly climbed into the model's top 20.

The Biggest Fallers: Who Got Exposed

Source: Boxscorus NCAAB BXS model, January 2026 game data.

Oklahoma and Auburn are the stories here. Both entered January as legitimate top-25 teams. Both left looking like bubble teams.

Oklahoma went 1-8 in the SEC. Their lone win came against Ole Miss on January 3rd (86-70). After that, eight consecutive losses:

Source: Boxscorus NCAAB game data, January 2026.

Four of those losses were by single digits. That's how the SEC gauntlet works. You don't have to be bad to go 1-8. You just have to be slightly below your conference's mean, and every team you play is good enough to punish that. Oklahoma entered January with a 1761 BXS that suggested they were an above-average SEC team. January corrected that assumption, to the tune of -52 points.

Auburn's story is different. They went 5-4, which doesn't look terrible. But the model weighs how you win and lose. Auburn lost to Georgia by 4 (100-104), Texas A&M by 2 (88-90), Missouri by 10 (74-84), and Tennessee by 8 (69-77). They beat Florida 76-67, which is impressive, but a team rated 1820 is supposed to be beating most SEC opponents by 5-10 points. Splitting nine games doesn't match a top-10 rating, and the model adjusted accordingly: -51 points.

The Big Ten Meat Grinder

So what happens when you put five of the country's top 11 teams in the same conference and let them beat each other up for a month?

Four of those top-11 Big Ten teams (Michigan at 1946, Illinois at 1881, Michigan State at 1880, and Nebraska at 1874) all gained BXS in January despite playing each other. How? They took it from the bottom of their own conference. Penn State dropped 66 BXS points. Oregon, USC, and Minnesota all declined. The Big Ten's top teams are cannibalizing their conference-mates.

Michigan went 8-1. Their only loss: Wisconsin beat them 91-88 on January 10th, when the model gave Michigan a 78.3% win probability (BXS: 1910 vs. 1752). I'll be honest, the model whiffed on that one. But Michigan bounced back and won their next seven, which is why their net gain for the month (+51) is still among the best in the country. That kind of resilience doesn't show up directly in BXS, but the results of resilience do.

Illinois went 8-0, quietly the most dominant January in the Big Ten. They beat Purdue 88-82 when Purdue was sitting at 1890 BXS, a top-5 team by the model's reckoning. They handled Penn State (73-65), Rutgers (81-55), Iowa (75-69), Northwestern (79-68), Minnesota (77-67), Maryland (89-70), and Washington (75-66) without a loss. Their +63 BXS gain (1818 to 1881) is one of the largest conference-play gains among any power-conference team this month.

The Purdue Question

Purdue is the team I keep coming back to. They entered January ranked 4th at 1865 BXS. They left at 1868. A gain of 2 points. Essentially flat.

But the trajectory hides a collapse. Purdue started 5-0, climbing to 1897 by January 17th. Then they lost three straight to close the month: UCLA (67-69), Illinois (82-88), and Indiana (67-72). The model sees a team that peaked mid-January and came back down. In a single month, Purdue went from 4th to 11th in the model's rankings without anyone really noticing.

So is Purdue still a contender? The model says "sure, but they stopped improving while everyone around them got better." That's the danger of going .500 in conference play: you don't drop off a cliff, but you watch the teams above you pull away. Every other serious contender gained meaningful BXS in January. Purdue gained 2 points.

Is 69.8% Accuracy Actually Good?

Across 1,558 January games, the model correctly picked the winner 69.8% of the time. My first reaction was mild embarrassment. Then I looked at the calibration data and realized the model is smarter than the headline number suggests.

Source: Boxscorus NCAAB BXS model, calibration data for January 2026.

The model said "this is a 50-60% game" 709 times, nearly half of all January games. In those games, the favorite lost 38.2% of the time. That's almost exactly what you'd expect. When the model said "70-80% favorite," the favorite won 81.7% of the time. When it said "80-90%," the favorite won 96.7%.

This is what good calibration looks like. The model's probabilities mean what they say. The 709 coin-flip games explain the "low" headline accuracy: the model correctly identified those games as unpredictable and didn't pretend otherwise. Anyone can predict Duke beating a 1300-rated team. The real test is predicting Michigan-Wisconsin, where the BXS gap is 158 points and home court advantage (65 BXS points in the Boxscorus model) closes that gap meaningfully. January is full of those tests.

For season-wide context, the model's accuracy through January is 73.2% across 5,856 games. The January-specific 69.8% is lower because conference games are between more evenly matched opponents. That's not a bug. That's the model being honest.

The Conference Power Rankings

Which conference is actually the toughest? The average BXS is a rough proxy: if you picked a random team from the conference and put them on a neutral court against a random team from another conference, the higher average wins more often.

Source: Boxscorus NCAAB BXS model, end of January 2026. Averages computed across all D1 members of each conference.

The SEC and Big Ten are separated by 0.6 BXS points. That's noise. The Big 12 sits 15 points behind, which is also negligible. The real gap is between the top three and everyone else: a 63-point drop to the ACC and Big East. For context, 63 BXS points translates to roughly a 59-41 edge on a neutral court. An average SEC team would be a solid favorite against an average ACC team.

What the averages hide: the Big Ten has the widest internal spread. Michigan (1946) to Penn State (1621) is a 325-point chasm. That's a conference where the best team would be an 85% favorite against the worst on a neutral court. The SEC and Big 12 are more compressed, which means more competitive games from top to bottom and fewer easy nights.

What January Tells You About February

It's January 31st. Conference tournaments are five weeks away. Selection Sunday is six weeks out. January gave the model its first real dataset, and the picture is already clearer than it was on New Year's Day.

A few principles from the data:

Trust the surgers. Arizona gained +65 in the Big 12, Duke gained +73 in the ACC, and Illinois gained +63 in the Big Ten. These aren't empty calories from beating sub-200 teams. These are verified gains against conference opponents the model rates highly. The model's opponent quality factor ensures that: a win against a 1700-rated team earns roughly 1.13x the normal K-factor, while a win against a 1300-rated team earns only 0.87x.

Watch the decliners. Oklahoma (-52) and Auburn (-51) are both SEC teams that entered January with top-25 BXS ratings and left looking like bubble teams. Their non-conference records overstated their quality. Kansas State (-54 in the Big 12) falls in the same category. February will tell us whether these drops are temporary slumps or permanent corrections.

Question the flat-liners. Purdue gained 2 BXS points in January. Iowa State gained +20 despite a 6-2 record, because their wins came by small margins. When every other contender is gaining 50+ points and your team is treading water, the model is telling you something.

February will add another 1,500+ conference games to the dataset. By March, the pretenders will have nowhere left to hide.

Check the full rankings at boxscorus.com/ncaab.